aiacademy: 深度學習貓狗大戰! model tuning

August 21, 2019 6 minute read

Tags: aiacademy, deep-learning, model-tuning, neural-networks, tensorflow

貓狗大戰

import

 import numpy as np 
 import pandas as pd
 import matplotlib.pyplot as plt
 import cv2
 import tensorflow as tf
 from tqdm import tqdm_notebook
 from sklearn.model_selection import train_test_split
 from sklearn.utils import shuffle

load data

 df = pd.read_csv('cat_dog_df.csv')
 df.head(10)
   
 pic = cv2.imread(df.file_name[0], cv2.IMREAD_GRAYSCALE)
 print(pic.shape)
 plt.imshow(pic, cmap='gray')
 plt.show()

split data adn preprocessing

 train_df, test_df = train_test_split(df, test_size=0.2, stratify=df.type)
   
 def df_to_data(df, picsize):
     img_ls = []
     label_ls = []
   
     for file_name, label in tqdm_notebook(zip(df.file_name, df.type)):
         img = cv2.imread(file_name, cv2.IMREAD_GRAYSCALE)
         img = cv2.resize(img, (picsize, picsize))
         img_ls.append(img)
           
         onehot = np.zeros(2)  
         onehot[label] = 1
         label_ls.append(onehot)
   
     x_data = np.array(img_ls).reshape(len(df), -1)  # flatten image data
     x_data = x_data / 255.  # normalization
     y_data = np.array(label_ls)
     return x_data, y_data

 picsize = 64
 X_train, y_train = df_to_data(train_df, picsize)
 X_test, y_test = df_to_data(test_df, picsize)
   
 print('size of training data:', X_train.shape, y_train.shape)
 print('size of testing data:', X_test.shape, y_test.shape)

build the network

 tf.reset_default_graph()
   
 with tf.name_scope('placeholder'):
     input_data = tf.placeholder(tf.float32, shape=[None, picsize*picsize], name='X')
     y_true = tf.placeholder(tf.float32, shape=[None, 2], name='y')
       
 with tf.variable_scope('network'):
     h1 = tf.layers.dense(input_data, 256, activation=tf.nn.sigmoid, name='hidden1')  # try to change the activation function
     h2 = tf.layers.dense(h1, 128, activation=tf.nn.sigmoid, name='hidden2') 
     h3 = tf.layers.dense(h2, 64, activation=tf.nn.sigmoid, name='hidden3')
     out = tf.layers.dense(h3, 2, name='output')
       
 with tf.name_scope('loss'):
     loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true, logits=out), name='loss')
       
 with tf.name_scope('accuracy'):
     correct_prediction = tf.equal(tf.argmax(tf.nn.softmax(out), 1), tf.argmax(y_true, 1))
     compute_acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
       
 with tf.name_scope('opt'):
     update = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss) # try to change the optimuizer and learning rate
       
 init = tf.global_variables_initializer()

show global variables

 tf.global_variables()
 # [<tf.Variable 'network/hidden1/kernel:0' shape=(4096, 256) dtype=float32_ref>,
 #  <tf.Variable 'network/hidden1/bias:0' shape=(256,) dtype=float32_ref>,
 #  <tf.Variable 'network/hidden2/kernel:0' shape=(256, 128) dtype=float32_ref>,
 #  <tf.Variable 'network/hidden2/bias:0' shape=(128,) dtype=float32_ref>,
 #  <tf.Variable 'network/hidden3/kernel:0' shape=(128, 64) dtype=float32_ref>,
 #  <tf.Variable 'network/hidden3/bias:0' shape=(64,) dtype=float32_ref>,
 #  <tf.Variable 'network/output/kernel:0' shape=(64, 2) dtype=float32_ref>,
 #  <tf.Variable 'network/output/bias:0' shape=(2,) dtype=float32_ref>]

算總共的weights!!!

自許為一個AI工程師要懂得如何計算出所有的weight唷!
畢卡葛手稿，幫助快速釐清

 var_sum = 0
 for tensor in tf.global_variables(scope='network'):
     if 'Adam' not in tensor.name:
         var_sum += np.product(tensor.shape.as_list())
       
 print('the total number of weights is', var_sum)  # 4096*256 + 256 + 256*128 + 128 + 128*64 + 64 + 64*2 + 2
 # the total number of weights is 1090114

訓練model啦!

 epoch = 100
 bs = 32
   
 train_loss_epoch, train_acc_epoch = [], []
 test_loss_epoch, test_acc_epoch = [], []
   
 sess = tf.Session()
 sess.run(init)
   
 for i in tqdm_notebook(range(epoch)):
       
 # training part
     train_loss_batch, train_acc_batch = [], []
       
     total_batch = len(X_train) // bs
       
     for j in range(total_batch):
           
         X_batch = X_train[j*bs : (j+1)*bs]
         y_batch = y_train[j*bs : (j+1)*bs]
         batch_loss, batch_acc, _ = sess.run([loss, compute_acc, update], 
                                             feed_dict={input_data: X_batch, y_true: y_batch})
           
         train_loss_batch.append(batch_loss)
         train_acc_batch.append(batch_acc)
           
     train_loss_epoch.append(np.mean(train_loss_batch))
     train_acc_epoch.append(np.mean(train_acc_batch))
       
 # testing part
     batch_loss, batch_acc = sess.run([loss, compute_acc], 
                                     feed_dict={input_data: X_test, y_true: y_test})
   
     test_loss_epoch.append(batch_loss)
     test_acc_epoch.append(batch_acc)
       
     X_train, y_train = shuffle(X_train, y_train)
       
     if i%5 == 0:
         print('step: {:2d}, train loss: {:.3f}, train acc: {:.3f}, test loss: {:.3f}, test acc: {:.3f}'
              .format(i, train_loss_epoch[i], train_acc_epoch[i], test_loss_epoch[i], test_acc_epoch[i]))
       

看訓練過程

 # step:  0, train loss: 0.742, train acc: 0.499, test loss: 0.693, test acc: 0.485
 # step:  5, train loss: 0.693, train acc: 0.507, test loss: 0.693, test acc: 0.546
 # step: 10, train loss: 0.692, train acc: 0.514, test loss: 0.693, test acc: 0.500
 # step: 15, train loss: 0.692, train acc: 0.529, test loss: 0.692, test acc: 0.559
 # step: 20, train loss: 0.692, train acc: 0.542, test loss: 0.692, test acc: 0.550
 # step: 25, train loss: 0.691, train acc: 0.542, test loss: 0.691, test acc: 0.555
 # step: 30, train loss: 0.691, train acc: 0.544, test loss: 0.691, test acc: 0.557
 # step: 35, train loss: 0.691, train acc: 0.552, test loss: 0.691, test acc: 0.502
 # step: 40, train loss: 0.690, train acc: 0.549, test loss: 0.690, test acc: 0.538
 # step: 45, train loss: 0.690, train acc: 0.551, test loss: 0.690, test acc: 0.554
 # step: 50, train loss: 0.690, train acc: 0.549, test loss: 0.690, test acc: 0.563
 # step: 55, train loss: 0.689, train acc: 0.552, test loss: 0.689, test acc: 0.561
 # step: 60, train loss: 0.689, train acc: 0.553, test loss: 0.689, test acc: 0.557
 # step: 65, train loss: 0.688, train acc: 0.555, test loss: 0.689, test acc: 0.562
 # step: 70, train loss: 0.688, train acc: 0.552, test loss: 0.688, test acc: 0.562
 # step: 75, train loss: 0.688, train acc: 0.555, test loss: 0.688, test acc: 0.563
 # step: 80, train loss: 0.687, train acc: 0.556, test loss: 0.688, test acc: 0.546
 # step: 85, train loss: 0.687, train acc: 0.559, test loss: 0.687, test acc: 0.562
 # step: 90, train loss: 0.687, train acc: 0.560, test loss: 0.687, test acc: 0.546
 # step: 95, train loss: 0.686, train acc: 0.560, test loss: 0.687, test acc: 0.558

close session
```
 sess.close()
```

出圖

 plt.plot(train_loss_epoch, 'b', label='train')
 plt.plot(test_loss_epoch, 'r', label='test')
 plt.legend()
 plt.title("Loss")
 plt.show()
   
 plt.plot(train_acc_epoch, 'b', label='train')
 plt.plot(test_acc_epoch, 'r', label='test')
 plt.legend()
 plt.title("Accuracy")
 plt.show()

先看結論

Loss & Accuracy
- 學太慢
- Accuracy太低:
  - train_set 和 test_set 的準確率 大約0.5! 我請我樓下三歲底笛來猜還比較準!!!

Model Tuning

改變 learning rate

學習速度太慢:

提升 learning rate: 改成 0.1 update = tf.train.GradientDescentOptimizer(learning_rate=0.1).minimize(loss)

 tf.reset_default_graph()
         
 with tf.name_scope('placeholder'):
     input_data = tf.placeholder(tf.float32, shape=[None, picsize*picsize], name='X')
     y_true = tf.placeholder(tf.float32, shape=[None, 2], name='y')
             
 with tf.variable_scope('network'):
     h1 = tf.layers.dense(input_data, 256, activation=tf.nn.sigmoid, name='hidden1')  # try to change the activation function
     h2 = tf.layers.dense(h1, 128, activation=tf.nn.sigmoid, name='hidden2') 
     h3 = tf.layers.dense(h2, 64, activation=tf.nn.sigmoid, name='hidden3')
     out = tf.layers.dense(h3, 2, name='output')
             
 with tf.name_scope('loss'):
     loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true, logits=out), name='loss')
             
 with tf.name_scope('accuracy'):
     correct_prediction = tf.equal(tf.argmax(tf.nn.softmax(out), 1), tf.argmax(y_true, 1))
     compute_acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
             
 with tf.name_scope('opt'):
     update = tf.train.GradientDescentOptimizer(learning_rate=0.1).minimize(loss) # try to change the optimuizer and learning rate
             
 init = tf.global_variables_initializer()

結果
- 學習速度變快
- training set 也進步到 0.85 正確率!!
- 有 overfitting 的情況發生!

Imgur

改變 activation function

sigmoid
- 把實數正無限大到負無限大映射到 1 ~ -1
- sigmoid 多次微分後會造成 vanishing gradient:
  - 改善方法: 換其他的　activation function

Relu

Imgur

大於零的資料: 不改變
小於零的資料: 回傳零
微分的狀態下:
- 大於零的資料: 1
- 小於零的資料: 0
Relu 特性資料通過不論幾次，樣子還是一樣

把activation function 換relu activation=tf.nn.relu，learning rate: 0.001

 tf.reset_default_graph()
         
 with tf.name_scope('placeholder'):
     input_data = tf.placeholder(tf.float32, shape=[None, picsize*picsize], name='X')
     y_true = tf.placeholder(tf.float32, shape=[None, 2], name='y')
             
 with tf.variable_scope('network'):
     h1 = tf.layers.dense(input_data, 256, activation=tf.nn.relu, name='hidden1')  # try to change the activation function
     h2 = tf.layers.dense(h1, 128, activation=tf.nn.relu, name='hidden2') 
     h3 = tf.layers.dense(h2, 64, activation=tf.nn.relu, name='hidden3')
     out = tf.layers.dense(h3, 2, name='output')
             
 with tf.name_scope('loss'):
     loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true, logits=out), name='loss')
             
 with tf.name_scope('accuracy'):
     correct_prediction = tf.equal(tf.argmax(tf.nn.softmax(out), 1), tf.argmax(y_true, 1))
     compute_acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
             
 with tf.name_scope('opt'):
     update = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss) # try to change the optimuizer and learning rate
             
 init = tf.global_variables_initializer()

結果
- 和改 learning rate 到 0.1 差不多
- train set 準確率有超過8成
- 還是有 overfitting 的情況發生!

改變優化器　optimizer

我的筆記

Adam

不會只計算現在的t度，而是會紀錄過去的t度為考量，而計算出更棒的兒數值

update = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)

 tf.reset_default_graph()
         
 with tf.name_scope('placeholder'):
     input_data = tf.placeholder(tf.float32, shape=[None, picsize*picsize], name='X')
     y_true = tf.placeholder(tf.float32, shape=[None, 2], name='y')
             
 with tf.variable_scope('network'):
     h1 = tf.layers.dense(input_data, 256, activation=tf.nn.relu, name='hidden1')  # try to change the activation function
     h2 = tf.layers.dense(h1, 128, activation=tf.nn.relu, name='hidden2') 
     h3 = tf.layers.dense(h2, 64, activation=tf.nn.relu, name='hidden3')
     out = tf.layers.dense(h3, 2, name='output')
             
 with tf.name_scope('loss'):
     loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true, logits=out), name='loss')
             
 with tf.name_scope('accuracy'):
     correct_prediction = tf.equal(tf.argmax(tf.nn.softmax(out), 1), tf.argmax(y_true, 1))
     compute_acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
             
 with tf.name_scope('opt'):
     update = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss) # try to change the optimuizer and learning rate
             
 init = tf.global_variables_initializer()

看tensor variables過去的紀錄

Imgur

計算 weight
- 會有三百多萬組，不是我們要的兒！

Imgur

加入 ‵if 'Adam' not in tensor.name:

  var_sum = 0
  for tensor in tf.global_variables(scope='network'):
      if 'Adam' not in tensor.name:
          var_sum += np.product(tensor.shape.as_list())
            
  print('the total number of weights is', var_sum)  # 4096*256 + 256 + 256*128 + 128 + 128*64 + 64 + 64*2 + 2
  # the total number of weights is 1090114

結果
- 7成多的成功率
- 還是有　overfitting 的狀況！

　　　　 Imgur

本機程式實作

先抓貓圖!

前行提要:
- 因為圖片在hub裡面想在本機完成的話要去，kaggle 下載貓狗檔案
- 助教給的csv /data/...，和我本機的位置不一樣
  所以偶去掉頭 / 寫 df.file_name[0][1:]

  import numpy as np
  import pandas as pd
  import matplotlib.pyplot as plt
  import cv2
  import tensorflow as tf
  from sklearn.model_selection import train_test_split
  from sklearn.utils import shuffle
    
    
  df = pd.read_csv('cat_dog_df.csv')
  print(df)
  # .str 可將 pandas 轉成string
  print(df['file_name'].str[1:])
    
  pic = cv2.imread(df.file_name[0][1:], cv2.IMREAD_GRAYSCALE)
  print(pic.shape)
  plt.imshow(pic, cmap='gray')
  plt.show()

Twitter Facebook LinkedIn

阿葛廷

aiacademy: 深度學習貓狗大戰! model tuning

貓狗大戰

先看結論

Model Tuning

本機程式實作

You May Also Enjoy

當你覺得對方很Ｇ8討厭的時候，放面鏡子照照自己剛剛的行為先吧！

daily Programming: 寶哥出場品質保證 GitHub Copilot

兩洞兩三最後三週計畫！

工作筆記: 想要開發然後開不了結果亂開亂關不熟的 K8S XDDD

阿葛廷

貓狗大戰

先看結論

Model Tuning

本機程式實作

You May Also Enjoy

當你覺得對方很Ｇ8討厭的時候，放面鏡子照照自己剛剛的行為先吧！

daily Programming: 寶哥出場 品質保證 GitHub Copilot

兩洞兩三最後三週計畫！

工作筆記: 想要開發然後開不了 結果亂開 亂關不熟的 K8S XDDD

daily Programming: 寶哥出場品質保證 GitHub Copilot

工作筆記: 想要開發然後開不了結果亂開亂關不熟的 K8S XDDD